Synchronization for OFDM-based over-the-air aggregation

Abstract

A system can receive respective data from respective sensors, wherein the respective data represents respective gradient values for a neural network produced by the respective first sensors according to a federated learning process. The system can transform the respective data into respective analog waveforms. The system can apply orthogonal frequency-division multiplexing to the respective analog waveforms to produce respective aligned analog waveforms. The system can create a superposition analog waveform that comprises a superposition of the respective aligned analog waveforms. The system can transmit the superposition analog waveform to an access point, wherein the access point is configured to update the neural network with the superposition analog waveform according to the federated learning process.

What is claimed is: 
     
       1. A system, comprising:
 a processor; and 
 a memory coupled to the processor, comprising instructions that, in response to execution by the processor, cause the system to perform operations, comprising:
 receiving respective data from respective sensors, wherein the respective data represents respective gradient values for a neural network produced by the respective first sensors according to a federated learning process; 
 transforming the respective data into respective analog waveforms; 
 applying orthogonal frequency-division multiplexing to the respective analog waveforms to produce respective aligned analog waveforms; 
 creating a superposition analog waveform that comprises a superposition of the respective aligned analog waveforms; and 
 transmitting the superposition analog waveform to an access point, wherein the access point is configured to update the neural network with the superposition analog waveform according to the federated learning process. 
 
 
     
     
       2. The system of  claim 1 , wherein applying orthogonal frequency-division multiplexing to the respective analog waveforms to produce the respective aligned analog waveforms comprises:
 applying respective time-domain cyclic prefixes to the respective analog waveforms. 
 
     
     
       3. The system of  claim 2 , wherein applying orthogonal frequency-division multiplexing to the respective analog waveforms to produce the respective aligned analog waveforms comprises:
 after applying the respective time-domain cyclic prefixes, aligning the respective analog waveforms in a frequency domain. 
 
     
     
       4. The system of  claim 2 , wherein applying orthogonal frequency-division multiplexing to the respective analog waveforms to produce the respective aligned analog waveforms comprises:
 determining a first length of the respective time-domain cyclic prefixes; and 
 determining a second length of an orthogonal frequency-division multiplexing symbol length, wherein the second length is greater than the first length by at least a scalar criterion. 
 
     
     
       5. The system of  claim 1 , wherein the operations further comprise:
 after applying orthogonal frequency-division multiplexing, correcting a phase noise effect of the respective aligned analog waveforms in a frequency domain. 
 
     
     
       6. The system of  claim 5 , wherein correcting the phase noise effect of the respective aligned analog waveforms in the frequency domain comprises:
 performing a signaling round before transmitting consecutive over-the-air frames for data transmission. 
 
     
     
       7. The system of  claim 1 , wherein the respective data is generated by the respective sensors by training respective local neural networks with respective local datasets. 
     
     
       8. The system of  claim 1 , wherein applying orthogonal frequency-division multiplexing to the respective analog waveforms to produce the respective aligned analog waveforms comprises:
 downsampling the respective analog waveforms to discrete values. 
 
     
     
       9. A method, comprising:
 receiving, by a system comprising a processor, respective data from respective sensors, wherein the respective data represents respective gradient values for a neural network; 
 transforming, by the system, the respective data into respective analog waveforms; 
 applying, by the system, orthogonal frequency-division multiplexing to the respective analog waveforms to produce respective aligned analog waveforms; 
 creating, by the system, a superposition analog waveform that comprises a superposition of the respective aligned analog waveforms; and 
 storing, by the system, the superposition analog waveform, wherein an access point is configured to update the neural network with the superposition analog waveform according to a federated learning process. 
 
     
     
       10. The method of  claim 9 , further comprising:
 sending, by the system, the superposition analog waveform to the access point. 
 
     
     
       11. The method of  claim 10 , wherein sending the superposition analog waveform to the access point comprises:
 sending an initialization preamble to the access point, wherein the initialization preamble comprises a frame timing subframe and a carrier frequency offset subframe. 
 
     
     
       12. The method of  claim 10 , wherein sending the superposition analog waveform to the access point comprises:
 sending a digital transmission frame to the access point, wherein the digital transmission frame comprises a frame timing subframe, a carrier frequency offset subframe, an orthogonal pilot sequence that comprises a first number of orthogonal frequency-division multiplexing symbols, and data symbols that comprise a second number of orthogonal frequency-division multiplexing symbols. 
 
     
     
       13. The method of  claim 12 , wherein the frame timing subframe comprises a third number of frame timing samples, and wherein the carrier frequency offset subframe comprises the third number of carrier frequency offset samples. 
     
     
       14. The method of  claim 10 , wherein sending the superposition analog waveform to the access point comprises:
 sending an over-the-air frame to the access point, wherein the over-the-air frame comprises a frame timing subframe, a common pilot subframe, and an over-the-air data sequence subframe. 
 
     
     
       15. The method of  claim 14 , wherein the frame timing subframe comprises a first number of frame timing samples, wherein the common pilot subframe comprises one orthogonal frequency-division multiplexing symbol, and wherein the over-the-air data sequence subframe comprises a second number of orthogonal frequency-division multiplexing symbols. 
     
     
       16. A non-transitory computer-readable medium comprising instructions that, in response to execution, cause a system comprising a processor to perform operations, comprising:
 receiving first data from a first sensor, wherein the first data represents first gradient values for a neural network; 
 receiving second data from a second sensor, wherein the second data represents second gradient values for the neural network; 
 transforming the first data into a first analog waveform; 
 transforming the second data into a second analog waveform; 
 applying orthogonal frequency-division multiplexing to the first analog waveform and the second analog waveform to produce a first aligned analog waveform, and a second aligned analog waveform, respectively; 
 creating a superposition analog waveform that comprises a superposition of the first aligned analog waveform and the second aligned analog waveform; and 
 storing the superposition analog waveform, wherein an access point is configured to update the neural network with the superposition analog waveform according to a federated learning process. 
 
     
     
       17. The non-transitory computer-readable medium of  claim 16 , wherein receiving the first data from the first sensor is based on the access point broadcasting a first digital transmission frame to trigger the first sensor and the second sensor, the first sensor starting a timer and recording a first current time based on receiving the first digital transmission frame, the first sensor feeding back a second digital transmission frame that is pre-equalized and recording a second current time, and wherein a phase error and a timing offset are determined by the access point based on receiving the second digital transmission frame. 
     
     
       18. The non-transitory computer-readable medium of  claim 17 , wherein receiving the first data from the first sensor is based on the access point broadcasting a third digital transmission frame to request the first data, the first sensor recording a third current time at which the third digital transmission frame is received, and the first sensor estimating an effective downlink channel. 
     
     
       19. The non-transitory computer-readable medium of  claim 18 , wherein receiving the first data from the first sensor is based on the first sensor modifying the first data based on the phase error and the timing offset. 
     
     
       20. The non-transitory computer-readable medium of  claim 19 , wherein the federated learning process comprises a machine-learning technique to train the neural network across multiple remote sensors that use respective local data, and independently of transmitting the respective local data.